Transition Policy

A reward for learning a transition policy is a success of following policy. A proximity prediction for proximity reward instead of binary reward estimates better on how close to good initial states. The paper defines proximity as $P(s) = \delta^{step}$ and $\delta \in (0,1)$ like discounting factor and provide proximity reward every step $P(s_{t+1}) - P(s_t)$. The paper learns proximity predictor by success buffer and failure buffer.